A Crowd Counting Algorithm Based On Deep Learning

With the development of China's transportation and the steady progress of urban construction,the population gradually gathered in cities,and the crowd gathering scene in public places such as stations,schools,squares has become increasingly prominent.It is necessary to monitor the dynamic change of the crowd effectively in these scenes.Among them,through statistical analysis of the number of people,we can alert the office for overcrowding,help the security department to formulate emergency plans and evacuate people,and prevent the occurrence of abnormal group events.According to the requirements of accuracy,real-time and robustness toward the current crowd counting task,a crowd counting algorithm is designed based on dense connection,and a crowd counting system is built using the proposed algorithm.The research contents are as follows:(1)The experiments analysis of the crowd counting algorithm based on YOLOv3 and the crowd counting algorithm based on CSRNet using density map prediction are carried out respectively.YOLOv3 algorithm is used to perform the head target detection.The network setting,prediction principle and training process are studied.The training and testing are carried out on the self-built data set,and the counting performance is tested.The test results show that YOLOv3 algorithm is easy to miss detection in dense crowd scenes,and the counting error is huge.The crowd count algorithm based on CSRNet is studied.The comparison experiment is designed for the dilated convolution network.The training and testing are carried out on Shanghai Tech A and Shanghai Tech B data sets.The experimental results show that the use of the dilated convolution does not significantly improve the performance of the network,but greatly slows the training speed of the model.(2)In view of the shortcomings of the existing crowd counting algorithms,a crowd counting algorithm is designed based on dense connection.The algorithm model includes two parts: front-end feature extraction network and back-end feature extraction network.In the front-end feature extraction network,a deeper convolution network is applied to increase the performance of the network;in the back-end feature extraction network,dense convolution block is designed to increase the transmission of the feature of the input between convolutional layers.In Shanghai Tech dataset and UCF?CC?50 dataset,the network is trained and tested.The experimental results show that the algorithm designed based on dense connection effectively reduces the counting error and increases the stability of the model.At the same time,compared with CSRNet algorithm,the training and testing time of the model is shortened,and the real-time requirement of the crowd counting task is met.(3)According to the algorithm of crowd counting based on dense connection,combined with the demand of crowd counting in dense crowd scenes,a crowd counting system is designed.The system can get the real-time crowd counting results by calling the counting model,analyze and compare the counting results,and give corresponding warning on the abnormal results.The test shows that the system can meet the needs of the normal crowd counting scene,and realize real-time functions,warning of massive people gathering in the scene.The counting method and counting system designed in this paper have good accuracy and calculation efficiency,and they have good practical value for density monitoring of crowded area.Also,it has a certain degree of reference significance for the academic research of crowd counting,station optimization,security monitoring and basic theory of artificial intelligence.